本文实现了一个两层的模拟异或(XOR)逻辑运算的神经网络。实现没有借助Pytorch等机器学习库,从而有助于读者更好地理解反向传播的计算细节。
我们知道感知学习(Perceptron learning)能表达的模型是线性的,如下图所示,其可以表达逻辑运算 AND, OR, NOR, 但是确不能表达 XOR。
但我们可以使用2层的网络,且引入能够让空间能够发生”扭曲“的激活函数(Activation Function),是能够让模型表达 XOR 的。
具体原理是 $x1\enspace XOR\enspace x2$ 可以被改写为: $(x_1\enspace AND\enspace x_2)\enspace NOR\enspace (x_1\enspace NOR\enspace x_2)$。如下图网络的结构所示,其中$x_1$, $x_2$是输入,$z$是输出,$g()$ 是激活函数。$u_1$ 和 $u_2$ 代表 $(x1 \enspace AND \enspace x2)$ 和 $(x1 \enspace NOR \enspace x2)$ 的运算结果:
我们可以手工推导可能的weights来验证模型的可行性:
# 选择 Step function 为激活函数 g:
def g(num):
if num >= 0:
return 1
else:
return 0
u1 = x1 AND x2 = x1 + x2 - 1.5 -> w11=1, w12=1, b1=-1.5
u2 = x1 NOR x2 = -x1 - x2 + 0.5 -> w21=-1, w22=-1, b2=0.5
s = y1 NOR y2 = -y1 - y2 + 0.5 -> v1=-1, v2=-1, c=0.5
s = -g(x1w11 + x2w12 + w10) - g(x1w21 + x2w22 + w20) + 0.5
= -g(x1 + x2 - 1.5) - g(-x1 - x2 + 0.5) + 0.5
# 验证:
x1=1, x2=1 -> s=-g(2 - 1.5) - g(-2 + 0.5) + 0.5 = -1 - 0 + 0.5 = -0.5 -> z=0
x1=0, x2=1 -> s=-g(1 - 1.5) - g(-1 + 0.5) + 0.5 = -0 - 0 + 0.5 = 0.5 -> z=1
x1=0, x2=0 -> s=-g(0 - 1.5) - g(0 + 0.5) + 0.5 = -0 - 1 + 0.5 = -0.5 -> z=0
延伸上面的逻辑改写的思路。我们知道任何的function都可以转化为一组DNF(Disjunctive Normal Form),这也就意味着任何的function都可以转化为2层的感知网络加activation。这也就不严谨的证明了万能近似定理。也就是说理论上2层的神经网络就可以表达任意的函数。
上述神经网络的求导过程入下:
∂E/∂z = z-t
∂z/∂s = z*(1-z)
∂E/∂s = ∂E/∂z * ∂z/∂s = (z-t) * z * (1-z) = δout
∂E/∂c = ∂E/∂s * ∂s/∂c = ∂E/∂s
∂E/∂v1 = δout * y1
∂E/∂v2 = δout * y2
∂E/∂y1 = δout * v1
∂E/∂y2 = δout * v2
∂y1/∂u1 = y1 *(1-y1)
∂E/∂u1 = ∂E/∂y1 * ∂y1/∂u1 = δout * v1 *y1 *(1-y1) = δ1
∂y2/∂u2 = y2 *(1-y2)
∂E/∂u2 = ∂E/∂y2 * ∂y2/∂u2 = δout * v2 *y2 *(1-y2)
∂u1/∂b1 = 1
∂u1/∂w11 = x1
∂u2/∂w21= x1
∂u1/∂w12 = x2
∂u2/∂w22 = x2
∂u2/∂b2 = 1
∂E/∂b1 = ∂E/∂u1 * ∂u1/∂b1
∂E/∂b2 = ∂E/∂u2 * ∂u2/∂b2
∂E/∂w11 = ∂E/∂u1 * ∂u1/∂w11 = δout * v1 *y1 *(1-y1) * x1 = (z-t) * z * (1-z) * v1 *y1 *(1-y1) * x1
- z = g( c + v1*g(b1 + w11x1 + w12x2) + v2*g(b2 + w21x1 + w22x2) )
- y1 = g(b1 + w11x1 + w12x2)
∂E/∂w12 = ∂E/∂u1 * ∂u1/∂w12 = δout * v1 *y1 *(1-y1) * x2
∂E/∂w21 = ∂E/∂u2 * ∂u2/∂w21 = δout * v2 *y2 *(1-y2) * x1
∂E/∂w22 = ∂E/∂u2 * ∂u2/∂w22 = δout * v2 *y2 *(1-y2) * x2
我们可以将上述求导过程实现为如下的代码。
该代码拷贝到本地后可以在CUP上进行训练。随机初始化的weights对是否能训练成功有一定影响,读者可以运行多次以进行尝试。
# learning XOR throw a two layer network use gradient descent
# w 的初始化值对是否能训练成功和训练的速度都有较大影响。这是无限模型带来的问题
import math
import random
import numpy as np
from datetime import datetime
# random.seed(3)
random.seed(datetime.now())
weightSet = set()
def random_weight():
while (True):
w = (random.randint(-10, 10)) / 10
if w == 0:
continue
if w not in weightSet:
weightSet.add(w)
return w
class Weights:
indexMap = {"b1": 0, "w11": 1, "w12": 2, "b2": 3, "w21": 4, "w22": 5, "c": 6, "v1": 7, "v2": 8}
weights = []
def __init__(self):
self.weights = [-1.5, 1, 1, 0.5, -1, -1, 0.5, -1, -1]
def getIndex(self, name):
return self.indexMap[name]
def get(self, index):
return self.weights[index]
def getWeights(self):
return self.weights
def update(self, index, w):
self.weights[index] = w
η = 0.05
weights = Weights()
for i in range(len(weights.getWeights())):
weights.update(i, random_weight())
# "b1", "w11", "w12", "b2", "w21", "w22", "c", "v1", "v2"
# weights.weights = [-1.5, 1, 1, 0.5, -1, -1, 0.5, -1, -1] # target
# target:
# weights.weights = [-3.6256294452213225, 2.472015878775767, 2.472015878775767, 1.4097549175948436, -3.0465380385550866,
# -3.0465380385550866, -3.7807463009027975, -4.221045218825591, -4.194984747863033]
# weights.weights = [0.5,-1,1,0.5,-1,-1,1.5,-1,-1]
# weights.weights = [-1.0, 0, 1, 0.5, -1, -1, 0.5, -1, -1]
# weights.weights = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0] # 不能初始化为0
inputs = [[1, 1], [1, 0], [0, 1], [0, 0]]
ts = [0, 1, 1, 0]
def cost():
return
λ = 0.001 # weight decay λ
def pW(weights, λ):
r = 0
for w in weights.weights:
r = r + w * w
r = (r * λ) / 2
return r
def sigmoid(num):
# return step(num)
# num = num * 5
r = 1 / (1 + math.exp(-num))
return r
def tanh(num) -> object:
r = 2 / (1 + math.exp(-2 * num)) - 1
return r
def step(num):
if num > 0:
return 1
else:
return 0
def activation(num):
# if(num)>0:
# return 1
# else:
# return 0
return sigmoid(num)
def calculate_s(inputs, weights):
result = weights[0]
for i in range(len(inputs)):
x = inputs[i]
result = result + x * weights[i + 1]
return result
def c_z(x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
return z
def c_z_step(x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = (x1 * w11 + x2 * w12 + b1)
if y1 > 0:
y1 = 1
else:
y1 = 0
y2 = (x1 * w21 + x2 * w22 + b2)
if y2 > 0:
y2 = 1
else:
y2 = 0
s = y1 * v1 + y2 * v2 + c
if s > 0:
s = 1
else:
s = 0
return s
def calculate_dE_dz(t, z, weights):
dE_dz = z - t
# dE_dz = dE_dz + λ * sum(weights.weights)
return dE_dz
def calculatePartialDerivatives_c(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dE_dc = dE_ds
return dE_dc
def calculatePartialDerivatives_v1(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dE_dv1 = dE_ds * y1
return dE_dv1
def calculatePartialDerivatives_v2(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dv2 = dout * y2
return dE_dv2
def calculatePartialDerivatives_b1(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dy1 = dout * v1
dy1_du1 = 1 - y1 ** 2
dE_du1 = dE_dy1 * dy1_du1
du1_db1 = 1
dE_db1 = dE_du1 * du1_db1
return dE_db1
def calculatePartialDerivatives_w11(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dy1 = dout * v1
dy1_du1 = 1 - y1 ** 2
dE_du1 = dE_dy1 * dy1_du1
du1_dw11 = x1
dE_dw11 = dE_du1 * du1_dw11
return dE_dw11
def calculatePartialDerivatives_w12(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dy1 = dout * v1
dy1_du1 = 1 - y1 ** 2
dE_du1 = dE_dy1 * dy1_du1
du1_dw12 = x2
dE_dw12 = dE_du1 * du1_dw12
return dE_dw12
def calculatePartialDerivatives_b2(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dy2 = dout * v2
dy2_du2 = 1 - y2 ** 2
dE_du2 = dE_dy2 * dy2_du2
du2_db2 = 1
dE_db2 = dE_du2 * du2_db2
return dE_db2
def calculatePartialDerivatives_w21(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dy2 = dout * v2
dy2_du2 = 1 - y2 ** 2
dE_du2 = dE_dy2 * dy2_du2
du2_dw21 = x1
dE_dw21 = dE_du2 * du2_dw21
return dE_dw21
def calculatePartialDerivatives_w22(t, x1, x2, weights):
b1, w11, w12, b2, w21, w22, c, v1, v2 = weights.getWeights()
y1 = tanh(x1 * w11 + x2 * w12 + b1)
y2 = tanh(x1 * w21 + x2 * w22 + b2)
s = y1 * v1 + y2 * v2 + c
z = activation(s)
dE_dz = calculate_dE_dz(t, z, weights)
dz_ds = z * (1 - z)
dE_ds = dE_dz * dz_ds
dout = dE_ds
dE_dy2 = dout * v2
dy2_du2 = 1 - y2 ** 2
dE_du2 = dE_dy2 * dy2_du2
du2_dw22 = x2
dE_dw22 = dE_du2 * du2_dw22
return dE_dw22
def calculatePartialDerivatives_sum(inputs, weights, calculatePartialDerivatives_single):
d = 0
for i in range(len(inputs)):
input = inputs[i]
t = ts[i]
x1 = input[0]
x2 = input[1]
d = d + calculatePartialDerivatives_single(t, x1, x2, weights)
# print(x1, x2, d)
return d
calulatePartialDerivativesFunctions = [
calculatePartialDerivatives_b1,
calculatePartialDerivatives_w11,
calculatePartialDerivatives_w12,
calculatePartialDerivatives_b2,
calculatePartialDerivatives_w21,
calculatePartialDerivatives_w22,
calculatePartialDerivatives_c,
calculatePartialDerivatives_v1,
calculatePartialDerivatives_v2]
def getCalulatePartialDerivativesFunction(index):
return calulatePartialDerivativesFunctions[index]
def calulateError(inputs, ts, weightIndex, w, weights):
weights.update(weightIndex, w)
# print(weights.__dict__)
E = 0
for i in range(len(inputs)):
input = inputs[i]
t = ts[i]
x1 = input[0]
x2 = input[1]
q = c_z(x1, x2, weights)
E = E + ((q - t) ** 2)
# p = activation(t)
# D_KL = (p * (math.log(p, 2) - math.log(q, 2)))
# E = E + D_KL
# print((z-t), (z-t)**2)
E = E / 2
# E = E + pW(weights, λ) # Weight decay
return E
epochs = 10000
# epochs = 10
def adjustWeights():
weightNames = ["b1", "w11", "w12", "b2", "w21", "w22", "c", "v1", "v2"]
# weightNames = ["w11", "w12"]
weightIndexs = []
for weightName in weightNames:
i = weights.getIndex(weightName)
weightIndexs.append(i)
derivatives = [0] * 9
for i in range(epochs):
# for i in range(1):
# 计算完所有的derivatives之后再更新ws或者每次都更新每个w都可以。
# 不能保证解决平原,马鞍等问题。但因为初始的w都比较小,从而很大概率?上避开了这些问题
for weightIndex in weightIndexs:
cF = getCalulatePartialDerivativesFunction(weightIndex)
d = calculatePartialDerivatives_sum(inputs, weights, cF)
derivatives[weightIndex] = d
# beforeW = weights.get(weightIndex)
# weights.update(weightIndex, beforeW - η * d)
for weightIndex in weightIndexs:
d = derivatives[weightIndex]
beforeW = weights.get(weightIndex)
weights.update(weightIndex, beforeW - η * d)
print("middle:", weights.__dict__)
def check():
E = 0
for i in range(len(inputs)):
input = inputs[i]
t = ts[i]
x1 = input[0]
x2 = input[1]
z = c_z(x1, x2, weights)
E = E + (z - t) ** 2
print("[{0},{1}] target: {2}, actual:{3}, E:{4}".format(x1, x2, t, z, E))
print("E", E / 2)
if __name__ == '__main__':
# weights.update(weights.getIndex("w21"), 10)
print("start:", weights.__dict__)
# check()
adjustWeights()
print("final:", weights.__dict__)
check()
训练输出如下:
...
final: {'weights': [-1.4450043568054496, 2.982519563105286, -3.0561807995146717, 1.365020614336376, 2.915482554021478, -2.835712113443007, 3.644913260490926, 4.032086569727394, -4.054900067525553]}
[1,1] target: 0, actual:0.025435505534571242, E:0.0006469649417992043
[1,0] target: 1, actual:0.9633013476840432, E:0.0019937560236066867
[0,1] target: 1, actual:0.9631037292967697, E:0.0033550908154127384
[0,0] target: 0, actual:0.028721019871054448, E:0.004179987797846243
E 0.0020899938989231213